Meta Deploys Millions of AMD EPYC Processors, AMD Invests $1 Billion in Meta

Meta deploys millions of EPYC processors, AMD invests another $1 billion Source: Ban Dao Zong Heng Published: 2026-10-28 13:57 Meta CPU Generate Poster From server CPUs to Instinct GPUs and Helios rack-scale systems, the two companies' collaboration is extending further into next-generation AI computing platforms and software ecosystems.

Recently, AMD announced a further deepening of its strategic partnership with Meta. Meta has already deployed millions of AMD EPYC server processors and has become a major customer for sixth-generation EPYC processors. The collaboration also covers Instinct GPUs, Helios rack-scale systems and software ecosystems, and will further extend to the Instinct MI500 series and next-generation rack-scale solutions in the future.

From server CPUs to AI accelerators and then to complete systems, the AMD-Meta collaboration is extending toward a more complete computing platform. In addition to hardware deployment, the two sides are also collaborating on model training, inference optimization and open-source software. In addition, AMD also announced that it will invest $1 billion in Meta's equity offering, further deepening the strategic ties between the two parties.

Meta becomes a major customer for sixth-generation EPYC

EPYC server processors are an important foundation of the long-term collaboration between AMD and Meta. AMD said Meta has deployed millions of EPYC processors across multiple generations in its infrastructure, and these CPUs also provide computing support for Meta's Muse personal AI agent.

The sixth-generation EPYC is codenamed Venice and uses the Zen 6 core architecture. For large-scale AI data centers, CPUs not only handle general-purpose computing tasks but are also responsible for system management, task scheduling and data processing, and provide support for GPU-accelerated computing. AI training and inference clusters require different components such as CPUs and GPUs to work together, and server processors remain an important part of the overall computing platform.

In February this year, AMD and Meta announced a multi-year, multi-generation product collaboration plan aimed at deploying up to 6 gigawatts of AMD Instinct GPUs. The first platforms will use Instinct GPUs customized based on the MI450 architecture, sixth-generation EPYC Venice processors, ROCm software and the Helios rack-scale architecture, with related products planned to begin shipping in the second half of 2026.

AMD also revealed that the two sides are collaborating on the development of subsequent EPYC processors, including Verano for future workloads. Meta becoming a major customer for the new generation of EPYC also means that its data center requirements can be integrated earlier into AMD's product planning, providing a foundation for subsequent deployments.

Helios and Instinct: collaboration deepens into AI computing systems

In addition to CPUs, AMD also hopes to take on more computing tasks in Meta's AI infrastructure through Instinct GPUs and Helios rack-scale systems.

This announcement confirms that the first AMD Helios rack-scale systems have already been installed and are being tested at Meta, with plans to further expand deployment in the future. Helios is aimed at large-scale AI computing, integrating compute chips, memory, networking and system design to provide rack-scale solutions.

For large AI clusters, GPU performance is only one of the factors affecting overall efficiency. Data transmission, memory access, network communication, power supply and cooling also affect system performance. By integrating CPUs, GPUs and rack-scale system design, AMD can perform collaborative optimization around customers' actual workloads.

On the GPU side, Meta has already been using AMD Instinct products in production environments for tasks such as model training, ranking and recommendation. Meta is also an early definition customer for the Instinct MI450 series and plans to deploy Instinct GPUs customized for its own workloads. In the future, the collaboration between the two sides will also expand to the Instinct MI500 series and next-generation rack-scale systems.

Among them, the MI350 series has already been used for relevant production workloads at Meta, MI450 is an important product in the two sides' large-scale deployment plan, and MI500 belongs to the direction of subsequent collaboration. This product roadmap shows that the collaboration between the two sides has extended from the application of existing products to the planning of next-generation computing platforms.

From chip deployment to software collaboration

AMD and Meta are optimizing recommendation model training for Instinct MI350X and contributing related improvements to open-source software ecosystems such as PyTorch, Triton and ROCm.

A technical paper published by Meta researchers uses AMD Instinct MI350X and NVIDIA B200 as examples to study how to optimize recommendation model training on different hardware platforms. Related research shows that directly migrating a model to another GPU architecture may not fully leverage hardware performance; it is also necessary to adjust the model structure, software configuration and low-level operators according to chip characteristics.

Through methods such as hardware-model co-design, performance analysis and kernel optimization, Meta increased end-to-end training throughput by up to about 1.5 times compared with the baseline solution of direct migration. The related improvements have been contributed to the PyTorch, Triton and ROCm open-source ecosystems.

In terms of model support, AMD has provided day-one support for Meta's Muse Spark Open, and developers can run the model on Instinct MI350X and MI355X GPUs through ROCm and vLLM. Muse Spark Open uses a mixture-of-experts model architecture, with a total parameter count of 816 billion, about 42 billion parameters activated per token, and support for a 1 million token context window.

Optimization data released by AMD further demonstrates the effect of software collaboration. On a single-node system equipped with 8 MI350X GPUs, after optimizations such as attention computation and mixture-of-experts modules, the MXFP4 version of the model increased output throughput from 206 tokens per second to 350 tokens per second under 4-way concurrent requests; under 32-way concurrent requests, it increased from 719 per second to 1,180 per second.

In addition to Muse Spark Open, the two sides have also promoted support for Muse Glimmer on AMD Ryzen AI personal computers and Radeon GPUs, extending model adaptation to more AMD hardware platforms.

AMD seeks a larger share of AI infrastructure

Meta currently uses NVIDIA GPUs, AMD GPUs and its self-developed MTIA accelerators simultaneously to meet the needs of different computing tasks. Parallel deployment of multiple computing platforms can increase the flexibility of hardware choices, but it also places higher demands on software migration and system optimization. Joint optimization between AMD and Meta helps reduce related adaptation costs.

From millions of EPYC processors to Instinct GPUs entering production environments, to Helios systems beginning testing, and to subsequent product plans such as MI500, the collaboration between AMD and Meta has covered multiple levels including chips, systems and software.

For AMD, a hyperscale customer like Meta can bring actual deployment demand, help verify product performance in complex production environments, and drive the improvement of the software ecosystem. For Meta, the collaboration among CPUs, GPUs and rack-scale systems, as well as coordination among different computing platforms, provides more choices for building AI infrastructure.

This article is reposted from media reports or online platforms and represents the author's personal stance or views. Our reposting is solely for sharing and does not mean that we endorse or agree with it. If the source attribution is incorrect or your lawful rights and interests are infringed, please contact customer service promptly, and as a neutral platform service provider we will promptly correct, delete or handle it in accordance with the law. Comments: No user comments yet. Latest News More Related Recommendations

LONGi Green Energy's Yinchuan Silicon Materials subsidiary increases capital to 3.4 billion yuan, an increase of 240% 2026-10-28 15:22

SmartSens' new subsidiary "Weiyun Chuangda" officially unveils three new AI computing power products 2026-10-28 15:21

The next decade of DRAM: 4F², 3D DRAM and cost projections for various roadmaps 2026-10-28 14:03